Corporate Earnings Calls and Analyst Beliefs

Battle Count: 93

The paper provides strong evidence that narrative content in earnings calls contains incremental predictive information beyond quantitative fundamentals, with statistically significant improvements in forecast accuracy (6-13% MSE reduction). The identification of systematic analyst biases (over-reaction to sentiment, under-reaction to uncertainty) directly informs potential trading strategies around earnings announcements. The PTE framework quantifies how specific narrative dimensions shift forecasts in basis points, enabling signal construction. However, the paper does not implement a full trading strategy with transaction costs, does not provide specific entry/exit rules, and the predictive gains, while statistically significant, are modest in absolute terms (1-6% R² improvement). The findings are most relevant for event-driven strategies around earnings calls and for understanding post-earnings-announcement drift (PEAD) mechanisms.

The methodology requires: (1) access to proprietary data (Capital IQ transcripts, IBES, CRSP, WRDS linking tables); (2) running FinBERT embeddings on 35,200+ transcripts with numeral masking; (3) training Gradient Boosting models with 446-1,214 features across multiple horizons; (4) generating counterfactual morphs using LLaMA 3 70B-Instruct (requiring significant GPU resources, as evidenced by CSCS supercomputing support); (5) implementing LLM-as-a-Judge validation pipeline; (6) computing PTEs across six narrative dimensions; (7) Clark and West statistical tests. The pipeline is multi-stage and computationally intensive, particularly the LLM-based morphing and validation steps. The masking procedure and embedding aggregation for long documents add complexity.

The paper provides detailed methodology descriptions including specific model identifiers (ProsusAI/finbert, LLaMA 3 70B-Instruct), prompt templates in appendices, data sources (Capital IQ Transcripts, IBES, CRSP, JKP dataset, Compustat), and hyperparameters (temperature=0.7, top_p=0.95). However, no code repository is mentioned, and the proprietary data sources (Capital IQ, IBES, WRDS) require institutional access. The LLM-as-a-Judge prompts are fully documented. The masking procedure and embedding methodology are described in detail.